Research On The Semi-supervised Learning Based On The Generative Adversarial Nets

The status of artificial intelligence has risen to the national strategic level,at the same time,as the core of artificial intelligence,machine learning undoubtedly has enormous development potential.It may even become the main driving force of the industrial revolution in the next phase.However,traditional supervised learning requires enough labeled data as a support for supervised learning algorithms,otherwise a supervised learning model with sufficient generalization performance cannot be obtained.Moreover,in industrial applications,the acquisition of the labeled data requires expert experience,which is time consuming and labor-intensive.Semi-supervised learning can achieve better results in practical applications,because a small amount of labeled data and a large amount of non-targeted data can be used for training.Unfortunately,most of the current research in semi-supervised learning focuses on solving the problem of lack of labeled samples based on the idea of label infiltration and data distribution models.The label infiltration model is usually obtain a learner using labeled data firstly,and optimize the learner by labeling unlabel data,such as Tri-Train etc.;The existing data distribution algorithm assumes that the samples follow a certain distribution and determine the model parameters by combining labeled and unlabeled data together.In order to solve the above problems,a semi-supervised learning model based on Generative Adversarial Nets is proposed in this paper.This method utilizes the fact that GAN can adaptively generate pseudo-samples similar to a given real sample,and provide training data effectively.This approach breaks the limitations of the original semi-supervised learning algorithm.Firstly,the initial KNN and SVM models are trained according to the labeled' data.Then the model s are used to mark the unlabeled data,and the results of the two models are comprehensively considered to select samples with high confidence and extend to the labeled data set.Then,the labeled samples that have been improved in terms of both quantity and quality are classified to form real sample data of different GANs respectively.So that the corresponding class samples are obtained through the GAN generator learning.Finally,all the samples are identified by classifiers and screened by two different types of learners.After obtaining high-quality training samples,the final classification model is trained.This method can effectively use the advantages of GAN to solve the problem of insufficient labeled data in semi-supervised learning by generating training data.Finally,the experimental results on the UCI benchmark proved that the algorithm is effective.